转载-原文地址：http://www.photoneray.com/opencl_04/
在前一篇介绍 program 等术语时，提到创建 program 对象有两种方式： clCreateProgramWithSource 和 clCreateProgramWithBinary。区别仅在于 opencl 程序在用户面前的展现形式，前者是源代码形式，后者是二进制形式。二进制形式的数据格式是不透明的，不同的实现可以有不同的标准。使用二进制形式的好处有二：一是由于二进制码已经经过编译（部分编译为中间件或全部编译为可执行文件），所以加载速度更快，需要的内存更少；二是可以保护 opencl 代码，保护知识产权。
下面我们就来看看如何利用二进制形式：
存储 opencl 程序为二进制
我们的第一个问题是：二进制版的 opencl 程序从哪里来？前文说过，所有的 cl 代码都要经过加载并创建 program 对象，然后由 program 对象在 device 上面编译并执行。难道还有其他方式编译 opencl 代码？答案是：NO!
意味着我们还是需要将代码送到 device 里面编译。你会说，这不是多此一举吗？看起来确实有点，不过一般的做法是在软件安装的时候就进行编译保存二进制形式，然后真正运行时才加载二进制。这样分成两个步骤的话，倒也说的过去。
省去前面那些与 platform、device 和 context的代码，我们直接进入创建 program 的地方。首先还是利用 clCreateProgramWithSource 函数读取源代码文件并用 clBuildProgram 函数编译。示例代码如下：
cl_int status;
cl_program program;
ifstream kernelFile("binary_kernel.ocl", ios::in);
if(!kernelFile.is_open())
return;
ostringstream oss;
oss << kernelFile.rdbuf();
string srcStdStr = oss.str();
const char *srcStr = srcStdStr.c_str();
program = clCreateProgramWithSource(context,
1,
(const char **)&srcStr,
NULL,
NULL);
if(program ==NULL)
return;
status = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
if(status != CL_SUCCESS)
return;
代码可能不完整，完整示例请看文末。
现在我们已经将 opencl 代码在 device 编译完成了。接下来要做的就是将编译好的二进制取出来存在磁盘上。使用的 API 就是 clGetProgramInfo:
cl_int clGetProgramInfo(cl_program program,
cl_program_info param_name,
size_t param_value_size,
void *param_value,
size_t *param_value_size_ret)
使用方法见如下代码片段（为使逻辑清晰，省略了错误检测，实际开发可不要省啊）：
cl_uint numDevices = 0;
// 获取 program 绑定过的 device 数量
clGetProgramInfo(program,
CL_PROGRAM_NUM_DEVICES,
sizeof(cl_uint),
&numDevices,
NULL);
// 获取所有的 device ID
cl_device_id *devices = new cl_device_id[numDevices];
clGetProgramInfo(program,
CL_PROGRAM_DEVICES,
sizeof(cl_device_id) * numDevices,
devices,
NULL);
// 决定每个 program 二进制的大小
size_t *programBinarySizes = new size_t[numDevices];
clGetProgramInfo(program,
CL_PROGRAM_BINARY_SIZES,
sizeof(size_t) * numDevices,
programBinarySizes,
NULL);
unsigned char **programBinaries = new unsigned char *[numDevices];
for(cl_uint i = 0; i < numDevices; ++i)
programBinaries[i] = new unsigned char[programBinarySizes[i]];
// 获取所有的 program 二进制
clGetProgramInfo(program,
CL_PROGRAM_BINARIES,
sizeof(unsigned char *) * numDevices,
programBinaries,
NULL);
// 存储 device 所需要的二进制
for(cl_uint i = 0; i < numDevices; ++i){
// 只存储 device 需要的二进制，多个 device 需要存储多个二进制
if(devices[i] == device){
FILE *fp = fopen("kernel_binary_ocl.bin", "wb");
fwrite(programBinaries[i], 1， programBinarySizes[i], fp);
fclose(fp);
break;
}
}
// 清理
delete[] devices;
delete [] programBinarySizes;
for(cl_uint i = 0; i < numDevices; ++i)
delete [] programBinaries[i];
delete[] programBinaries;
要注意的是，可能有很多个 device 都编译了 program，所以将二进制提取出来时，我们是遍历了所有编译了 program 的 device。
读取二进制版opencl程序
经过上面一系列的操作，我们的磁盘上应该存在一个二进制版的 opencl 程序了。里面的内容可能是可读的，也可能是不可读的。这个视不同厂商实现而不同。
相对于存储，读取看起来就清爽的多，无非是打开二进制文件，然后调用 clCreateProgramWithBinary函数。示例如下：
FILE *fp= fopen("kernel_binary_ocl.bin", "rb");
// 获取二进制的大小
size_t binarySize;
fseek(fp, 0, SEEK_END);
binarySize = ftell(fp);
rewind(fp);
// 加载二进制文件
unsigned char *programBinary = new unsigned char[binarySize];
fread(programBinary, 1, binarySize, fp);
fclose(fp);
cl_program program;
program = clCreateProgramWithBinary(context,
1,
&device,
&binarySize,
(const unsigned char**)&programBinary,
NULL，
NULL）；
delete [] programBinary;
clBuildProgram(program 0, NULL, NULL, NULL, NULL);
这里要注意，即使加载是二进制版，我们在之后还是要对其进行 clBuildProgram。原因在于，我们无法保证所谓的二进制一定是可执行码。因为每个厂商的实现不一，有的可能就是最终执行码，而有的却是中间码。所以无论是从源代码还是二进制创建 program，之后都需要 clBuildProgram。
这样兜了一圈，发现要使用二进制版还是用了一遍源代码方式，感觉代码复杂好多，有点多余。其实换个角度来看，我们完全可以写成两个程序，一个专门用来读取源代码并编译生成二进制，另一个才是读取二进制运行软件。前者开发人员使用，后者才是给用户使用的。只有这样才能体现二进制版的优势。
示例代码
OpenCLCompileToBin.cpp
/* OpenCLCompileToBin.cpp
* (c) by keyring <keyrings@163.com>
* 2016.01.13
*/
#include <iostream>
#if defined(__APPLE__) || defined(__MACOSX)
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
#define KERNEL(...)#__VA_ARGS__
const char *kernelSourceCode = KERNEL(
__kernel void hellocl(__global uint *buffer)
{
size_t gidx = get_global_id(0);
size_t gidy = get_global_id(1);
size_t lidx = get_local_id(0);
buffer[gidx + 4 * gidy] = (1 << gidx) | (0x10 << gidy);
}
);
int main(int argc, char const *argv[])
{
printf("hello OpenCL\n");
cl_int status = 0;
size_t deviceListSize;
// 得到并选择可用平台
cl_uint numPlatforms;
cl_platform_id platform = NULL;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
if (status != CL_SUCCESS) {
printf("ERROR: Getting Platforms.(clGetPlatformIDs)\n");
return EXIT_FAILURE;
}
if (numPlatforms > 0) {
cl_platform_id *platforms = (cl_platform_id *)malloc(numPlatforms * sizeof(cl_platform_id));
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
if (status != CL_SUCCESS) {
printf("Error: Getting Platform Ids.(clGetPlatformIDs)\n");
return -1;
}
for (unsigned int i = 0; i < numPlatforms; ++i) {
char pbuff[100];
status = clGetPlatformInfo(
platforms[i],
CL_PLATFORM_VENDOR,
sizeof(pbuff),
pbuff,
NULL);
platform = platforms[i];
if (!strcmp(pbuff, "Advanced Micro Devices, Inc.")) {
break;
}
}
delete platforms;
}
// 如果我们能找到相应平台，就使用它，否则返回NULL
cl_context_properties cps[3] = {
CL_CONTEXT_PLATFORM,
(cl_context_properties)platform,
0
};
cl_context_properties *cprops = (NULL == platform) ? NULL : cps;
// 生成 context
cl_context context = clCreateContextFromType(
cprops,
CL_DEVICE_TYPE_GPU,
NULL,
NULL,
&status);
if (status != CL_SUCCESS) {
printf("Error: Creating Context.(clCreateContexFromType)\n");
return EXIT_FAILURE;
}
// 寻找OpenCL设备
// 首先得到设备列表的长度
status = clGetContextInfo(context,
CL_CONTEXT_DEVICES,
0,
NULL,
&deviceListSize);
if (status != CL_SUCCESS) {
printf("Error: Getting Context Info device list size, clGetContextInfo)\n");
return EXIT_FAILURE;
}
cl_device_id *devices = (cl_device_id *)malloc(deviceListSize);
if (devices == 0) {
printf("Error: No devices found.\n");
return EXIT_FAILURE;
}
// 现在得到设备列表
status = clGetContextInfo(context,
CL_CONTEXT_DEVICES,
deviceListSize,
devices,
NULL);
if (status != CL_SUCCESS) {
printf("Error: Getting Context Info (device list, clGetContextInfo)\n");
return EXIT_FAILURE;
}
// 装载内核程序，编译CL program ,生成CL内核实例
size_t sourceSize[] = {strlen(kernelSourceCode)};
cl_program program = clCreateProgramWithSource(context,
1,
&kernelSourceCode,
sourceSize,
&status);
if (status != CL_SUCCESS) {
printf("Error: Loading Binary into cl_program (clCreateProgramWithBinary)\n");
return EXIT_FAILURE;
}
// 为指定的设备编译CL program.
status = clBuildProgram(program, 1, devices, NULL, NULL, NULL);
if (status != CL_SUCCESS) {
printf("Error: Building Program (clBuildingProgram)\n");
return EXIT_FAILURE;
}
deviceListSize = 1;
// 决定每个 program 二进制的大小
size_t *programBinarySizes = new size_t[deviceListSize];
clGetProgramInfo(program,
CL_PROGRAM_BINARY_SIZES,
sizeof(size_t) * deviceListSize,
programBinarySizes,
NULL);
printf("%lu\n", deviceListSize);
unsigned char **programBinaries = new unsigned char *[deviceListSize];
for(cl_uint i = 0; i < deviceListSize; ++i)
programBinaries[i] = new unsigned char[programBinarySizes[i]];
// 获取所有的 program 二进制
clGetProgramInfo(program,
CL_PROGRAM_BINARIES,
sizeof(unsigned char *) * deviceListSize,
programBinaries,
NULL);
printf("ready write to file\n");
// 写入文件
FILE *fp = fopen("./kernel_binary_ocl.bin", "wb");
fwrite(programBinaries[0], 1, programBinarySizes[0], fp);
fclose(fp);
// 资源回收
status = clReleaseProgram(program);
status = clReleaseContext(context);
free(devices);
delete [] programBinarySizes;
for(cl_uint i = 0; i < deviceListSize; ++i)
delete [] programBinaries[i];
delete programBinaries;
return 0;
}
OpenCLRunWithBin.cpp
/* OpenCLRunWithBin.cpp
* (c) by keyring <keyrings@163.com>
* 2016.01.13
*/
#include <iostream>
#if defined(__APPLE__) || defined(__MACOSX)
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
// #define KERNEL(...)#__VA_ARGS__
// const char *kernelSourceCode = KERNEL(
// __kernel void hellocl(__global uint *buffer)
// {
// size_t gidx = get_global_id(0);
// size_t gidy = get_global_id(1);
// size_t lidx = get_local_id(0);
// buffer[gidx + 4 * gidy] = (1 << gidx) | (0x10 << gidy);
// }
// );
int main(int argc, char const *argv[])
{
printf("hello OpenCL\n");
cl_int status = 0;
size_t deviceListSize;
// 得到并选择可用平台
cl_uint numPlatforms;
cl_platform_id platform = NULL;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
if (status != CL_SUCCESS) {
printf("ERROR: Getting Platforms.(clGetPlatformIDs)\n");
return EXIT_FAILURE;
}
if (numPlatforms > 0) {
cl_platform_id *platforms = (cl_platform_id *)malloc(numPlatforms * sizeof(cl_platform_id));
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
if (status != CL_SUCCESS) {
printf("Error: Getting Platform Ids.(clGetPlatformIDs)\n");
return -1;
}
for (unsigned int i = 0; i < numPlatforms; ++i) {
char pbuff[100];
status = clGetPlatformInfo(
platforms[i],
CL_PLATFORM_VENDOR,
sizeof(pbuff),
pbuff,
NULL);
platform = platforms[i];
if (!strcmp(pbuff, "Advanced Micro Devices, Inc.")) {
break;
}
}
delete platforms;
}
// 如果我们能找到相应平台，就使用它，否则返回NULL
cl_context_properties cps[3] = {
CL_CONTEXT_PLATFORM,
(cl_context_properties)platform,
0
};
cl_context_properties *cprops = (NULL == platform) ? NULL : cps;
// 生成 context
cl_context context = clCreateContextFromType(
cprops,
CL_DEVICE_TYPE_GPU,
NULL,
NULL,
&status);
if (status != CL_SUCCESS) {
printf("Error: Creating Context.(clCreateContexFromType)\n");
return EXIT_FAILURE;
}
// 寻找OpenCL设备
// 首先得到设备列表的长度
status = clGetContextInfo(context,
CL_CONTEXT_DEVICES,
0,
NULL,
&deviceListSize);
if (status != CL_SUCCESS) {
printf("Error: Getting Context Info device list size, clGetContextInfo)\n");
return EXIT_FAILURE;
}
cl_device_id *devices = (cl_device_id *)malloc(deviceListSize);
if (devices == 0) {
printf("Error: No devices found.\n");
return EXIT_FAILURE;
}
// 现在得到设备列表
status = clGetContextInfo(context,
CL_CONTEXT_DEVICES,
deviceListSize,
devices,
NULL);
if (status != CL_SUCCESS) {
printf("Error: Getting Context Info (device list, clGetContextInfo)\n");
return EXIT_FAILURE;
}
FILE *fp= fopen("./kernel_binary_ocl.bin", "rb");
// 获取二进制的大小
size_t binarySize;
fseek(fp, 0, SEEK_END);
binarySize = ftell(fp);
rewind(fp);
// 加载二进制文件
unsigned char *programBinary = new unsigned char[binarySize];
fread(programBinary, 1, binarySize, fp);
fclose(fp);
cl_program program;
program = clCreateProgramWithBinary(context,
1,
&devices[0],
&binarySize,
(const unsigned char**)&programBinary,
NULL,
NULL);
delete [] programBinary;
// 装载内核程序，编译CL program ,生成CL内核实例
// size_t sourceSize[] = {strlen(kernelSourceCode)};
// cl_program program = clCreateProgramWithSource(context,
// 1,
// &kernelSourceCode,
// sourceSize,
// &status);
// if (status != CL_SUCCESS) {
// printf("Error: Loading Binary into cl_program (clCreateProgramWithBinary)\n");
// return EXIT_FAILURE;
// }
// 为指定的设备编译CL program.
status = clBuildProgram(program, 1, devices, NULL, NULL, NULL);
if (status != CL_SUCCESS) {
printf("Error: Building Program (clBuildingProgram)\n");
return EXIT_FAILURE;
}
// 得到指定名字的内核实例的句柄
cl_kernel kernel = clCreateKernel(program, "hellocl", &status);
if (status != CL_SUCCESS) {
printf("Error: Creating Kernel from program.(clCreateKernel)\n");
return EXIT_FAILURE;
}
// 创建 OpenCL buffer 对象
unsigned int *outbuffer = new unsigned int [4 * 4];
memset(outbuffer, 0, 4 * 4 * 4);
cl_mem outputBuffer = clCreateBuffer(
context,
CL_MEM_ALLOC_HOST_PTR,
4 * 4 * 4,
NULL,
&status);
if (status != CL_SUCCESS) {
printf("Error: Create Buffer, outputBuffer. (clCreateBuffer)\n");
return EXIT_FAILURE;
}
// 为内核程序设置参数
status = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&outputBuffer);
if (status != CL_SUCCESS) {
printf("Error: Setting kernel argument. (clSetKernelArg)\n");
return EXIT_FAILURE;
}
// 创建一个OpenCL command queue
cl_command_queue commandQueue = clCreateCommandQueue(context,
devices[0],
0,
&status);
if (status != CL_SUCCESS) {
printf("Error: Create Command Queue. (clCreateCommandQueue)\n");
return EXIT_FAILURE;
}
// 将一个kernel 放入 command queue
size_t globalThreads[] = {4, 4};
size_t localThreads[] = {2, 2};
status = clEnqueueNDRangeKernel(commandQueue, kernel,
2, NULL, globalThreads,
localThreads, 0,
NULL, NULL);
if (status != CL_SUCCESS) {
printf("Error: Enqueueing kernel\n");
return EXIT_FAILURE;
}
// 确认 command queue 中所有命令都执行完毕
status = clFinish(commandQueue);
if (status != CL_SUCCESS) {
printf("Error: Finish command queue\n");
return EXIT_FAILURE;
}
// 将内存对象中的结果读回Host
status = clEnqueueReadBuffer(commandQueue,
outputBuffer, CL_TRUE, 0,
4 * 4 * 4, outbuffer, 0, NULL, NULL);
if (status != CL_SUCCESS) {
printf("Error: Read buffer queue\n");
return EXIT_FAILURE;
}
// Host端打印结果
printf("out:\n");
for (int i = 0; i < 16; ++i) {
printf("%x ", outbuffer[i]);
if ((i + 1) % 4 == 0)
printf("\n");
}
// 资源回收
status = clReleaseKernel(kernel);
status = clReleaseProgram(program);
status = clReleaseMemObject(outputBuffer);
status = clReleaseCommandQueue(commandQueue);
status = clReleaseContext(context);
free(devices);
delete outbuffer;
system("pause");
return 0;
}
先使用compile文件编译一个bin，然后使用run文件加载bin运行。
扫一扫
538
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
